EVTA Seminar

Vrije University Amsterdam (October 8, 2026)

Location: room Spectrum 5, VU Onderzoeksgebouw VU, VU Amsterdam

Address: De Boelelaan 1100, 1081 HZ Amsterdam

12:00-13:00 Olivier Wintenberger (Sorbonne Université)

Title: RBM procedure under Pareto or Fréchet second order regular variation

Abstract: Under the Pareto second order regular variation condition, Wager (2014) proved the asymptotic normality of the Random Block Maxima (RBM) estimator. This U-statistic can be easily adapted to improve the convergence rate under the Fréchet second order regular variation condition. We propose a new methodology based on both RBM estimators to select adaptively the tuning parameter and to debias estimates by leveraging the two conditions simultaneously, whichever holds being unknown a priori. To do so, one has to robustly estimate the rate of convergence from the second order conditions without estimating the second order parameter (based on Hall (1990)'s procedure). Then we apply a debiasing procedure similar to Richardson’s; See Bach (2021) for instance, and finally combine both RBMs to obtain a better variance than the best of the two taken individually. An empirical study illustrates the effectiveness of the procedure, especially for reducing the bias.

This is a joint work with Gloria Buriticá (MIA Paris-Saclay, AgroParisTech) and Thomas Mikosch (Department of Mathematics, Copenhagen University).

13:00-13:30 Lunch

13:30-14:15 John Einmahl (Tilburg University)

Title: Refining disaster risk assessment through heterogeneity inference 

Abstract: We consider univariate heterogeneous data in a heavy-tailed setting. The asymptotic variance of the Hill estimator and of extreme quantile estimators is known to be equal to that in the i.i.d. case times a variance reduction factor that depends on the degree of heterogeneity. For statistical inference this factor has to be estimated, but a feasible consistent estimator is not known yet. We first provide two estimators of the reduction factor, show their consistency, and assess by simulations how the resulting variance estimators affect finite-sample inference for the extreme value index and extreme quantiles. Second, our novel procedures are particularly relevant for disaster risk assessment, where losses arise from different types of events, locations, and exposure levels. This heterogeneity makes reliable uncertainty quantification challenging, but it is now within reach since we can estimate the variances consistently. We provide an application to natural mass disasters, including floods, storms, extreme temperatures, wildfires, earthquakes, and other geophysical and hydrometeorological hazards.

This is a Joint work with Abdelaati Daouia (Toulouse) and Gilles Stupfler (Angers).